Adjustment with least squares method

Two software packages – Two results

Petra Radomi, Markus Schlösser
DESY Hamburg

Adjustment with least Design of a simulated 3-D-network

squares method Development Adjustment Upgrading the 3D-network to 5 instruments

A 3D-network with 10 points and three instruments was simulated, The output of PANDA includes the adjusted coordinates and Because of the good corresponding results in this network with three
covering an area of 30 * 30 m² with elevation difference up to 4m. observations, also the correction of the observations and the sum of instruments, two more instruments (dark blue in figure 2) were
Several software packages use the “least squares method” for squares of the corrections.
Because of the fact, that in SA an instrument stand cannot be a integrated and the number of observations rises up to 135. These
adjustment of observations to calculate the coordinates. The observations were also distorted, but with new random numbers, and
measured point, the instruments have the same theoretical
mathematical background of this method is to minimize the weighted SA represents the adjusted coordinates of the measured points in a
coordinates as the identical target points but a different name. again the observations were adjusted. In the figure below the vectors
sum of squares,v Pv → min consequently the evaluation of this sum
T

For all network simulations the observation data is related to gravity. point group, the coordinates of the instruments are shown in the of correction are shown:
is a possibility to compare the results of different programs and their instrument properties.
algorithms.
Vectors of correction: Magnification 100000 PANDA black SA red
Based on a simulated network the true coordinates are known. The differences to the true coordinates are shown in figure 3.
zzzz

zzzz

X = [x T
y . . . x y ] Netw ork design w ith 3 and 5 instrum ents
1111

nnnn

0 1 1 n n
Vectors of correction: Magnification 100000 PANDA black SA red
20
The true observations are calculated from the true coordinates.
15

LT = [r . d]
~
1 . r n z 1 . z n d 1 n 10

The standard deviation of azimuth and zenith is set to σ = 0.3mgon, r / z

5

the standard deviation of distance is set to σ =0.05mm+0ppm. d

- 20 -15 -10 -5
0

-5 0 5 10 15 20

~ -10
The observation vector L consists of the true observations L and a
random error vector ε , which is scaled with the standard deviation of
-15

-20
each obervation.
~ t ar get point s inst r ument st ands addit ional inst r ument st ands
L = L +ε
Figure 2

εi = t × σ i r / z / d

ε T = [εr . εr 1 n εz 1 . εz n εd 1 . εd ] n
Calculation Figure 5

The ‚true‘ observations (distance, azimuth, zenit) for the three Figure 4
The Gaussian variables t were randomized with the Box-Muller-i
instruments were calculated. Sum of squares
Method. Such numbers can be numerically sampled from two uniform
([0, 1]) random numbers u1 and u2 through the formula
These 81 observations were distorted by the described method and Result
imported in PANDA and Spatial Analyzer. The true point-coordinates distance azimuth zenith Sum of
t = cos(2 * π * u ) * ( −2 * ln u ) 1 2
were imported as coarse coordinates to locate the instruments in SA The adjusted coordinates are used to calculate the adjusted squares
and for further calculations in the software packages. observations in order to get the corrections and their sum of squares. PANDA 27,02930 20,36210 33,87020 81,26160
Normal Distribution
3000 numbers randomized with Box-Muller-method PANDA calc. 28,26602 20,53047 34,01268 82,80916
The standard deviations were set to 0.3mgon and 0.05mm+0ppm. This step is not really necessary in PANDA, but was done to get a
PANDA runs a so called ‘free adjustment‘ no point is fixed, all points feeling for the accuracy of calculation. SA –USMN calc. 28,00942 28,95287 33,60847 90,57076
300
together define the datum of the network. Table 2
250 The sums of squares are shown in the following table:
In SA a so called ‚Unified Spatial Metrology Network‘ is calculated,
200
rotations around x and y are disabled for all instruments. The adjusted
Two more instrument stands and 54 more observations have of
150 coordinates are saved in a new point group and the instruments are distance azimuth zenith Sum of
squares course an effect on the individual sum of squares for distance and
transformed in SA.
100
zenith, but not really for the azimuth. In the azimuth sum of squares
PANDA 6,24610 18,10920 27,56550 51,92080
of Spatial Analyzer there is even a small reduction.
50 PANDA calc. 6,29878 17,93526 27,62089 51,85494
SA –USMN calc. 6,20966 29,14863 27,35207 62,71036
0
-4 -3 -2 -1 0 1 2 3 4 SA-USMN with point 7,11357 26,73884 28,87664 62,72905

Figure 1 Table 1

Summary
As shown in the table for this network a ‘Unified Spatial Metrology
Now the generated observations can be adjusted considering the Network with point group’ was calculated,too. The orignal coordinates Surprisingly the sum of squares of Spatial Analyzer and PANDA are
least squares method were allowed to move. very similar concerning the distance and zenith angle, but for the
azimuth Spatial Analyzer gives a much larger value than PANDA.
v T Pv → min
∧
There are little variances between USMN with and without point
The adjusted observations L are the sum of the observations L groups in the individual sum of squares, but finally the total sum of Therefore, from the analytical point of view the Spatial Analyzer result
and the correction v . squares doesn‘t change. is suboptimal.
∧
L = L +v

In ideal case the correction v should be equal to the random error ε

Figure 3

Design of a 3-D Linear-Accelerator-Network

Network configuration Result Redundancy Measurement setup

Additionally, a second network has been used for comparison. Vectors of correction: Magnification 100000 PANDA black SA red To answer the question whether the difference at the end of the Last but not least: It is particularly remarkable that the SA vectors of
It has been simulated as a linear accelerator network, which is similar network can be explained with the unused observations, the network correction grow from top to bottom or from instrument 1 to instrument
to the ones commonly used at DESY. was extended with two additional instruments so that now the last ring 9 whereas the size of the PANDA vectors of correction is consistent in
is also redundant. this net.
Four points, left and right of the theoretical beamline, two of them on
the bottom, the other on the ceiling, build a ‚ring‘. Every ten meters Now nine instruments are used, the number of (new) distorted So, does the solution in SA depend on the internal order of the
another ring is installed. Overall there are nine rings. The instrument observations is 468, while the number of target points is still 36. instruments?
is placed in the middle of two rings and observes usually two rings There are two possibilities to check this idea.
forward and two rings backward.
Vectors of correction: Magnification 100000 PANDA black SA red First one is to copy the instruments in reverse order in a new
In this example seven instruments are used, so that the last ring is collection and adjust the network again. There is no difference in the
measured without redundancy. results for this approach.

The second possibility is to create an new collection and read in the

instruments and point groups (measurements) in reverse order. The
adjustment of this data shows a mirrored result for the vectors of
Linear network design with 7 and 9 instruments
correction.
4
Figure 7
That means that the timestamp of the measurements is a critical
3
information for the adjustment of Spatial Analyzer.
2 While for the PANDA solution the distribution of correction looks
1 relatively consistent, for the SA result a lateral movement of the Vectors of correction: Magnification 100000 SA red
0
adjusted coordinates is pronounced. A short look at the sum of
-5 5 15 25 35 45 55 65 75 85 95 squares for azimuths confirms that the SA-solution is 1.5 times
-1
higher than the PANDA result for the azimuths.
-2

-3 In this network the individual sum of squares for distance and Figure 8
Zenith are equal for both results within the limits of accuracy of
-4
calculation. distance azimuth zenith Sum of
target points instrument stands additional instrument stands squares
However, the total sum of squares is significantly larger for SA than PANDA 117,83290 94,58250 104,68590 317,10130
Figure 6 for PANDA.
PANDA calc. 117,07540 94,89817 104,82915 316,80271
SA –USMN calc. 134,63800 122,82293 122,49341 379,95434
distance azimuth zenith Sum of
squares Table 4
Calculation
PANDA 98,68740 76,24530 85,48060 260,41330
The result of this linear network is again surprising: the sum of
The ‚true‘ observations (distance, azimuth, zenith) for the seven PANDA calc. 96,48888 76,18897 85,26244 257,94028 squares for azimuths of the Spatial Analyzer solution doesn‘t really
instruments and 36 measured points were calculated. The 384 SA –USMN calc. 97,07441 122,50218 85,44770 305,02429 change, but now there are also differences to the PANDA adjustment
observations were distorted and imported in PANDA and Spatial in the other sums of squares, which didn‘t exist in the other networks.
Figure 9
Analyzer. Table 3

PANDA adjusted the 384 observations and didn‘t show corrections for
the 12 observations to the four points of the last ring, which are only
observed once (as said before).
Conclusion
SA adjusted only the 372 observations to the points which are at least A method for the comparison of different network adjustment packages With the comparison of the sum of squares obviously the rightness of Especially the lateral movement of points from start to end is important
observed twice, but surprisingly the 12 unchecked observations also has been shown. So far only PANDA and SA have been evaluated with a solution can not be proven. So at the moment the minimal sum of for the acclerator community. This behaviour ist dependent on the
changed. At the moment there is no explanation for this behaviour. a limited set of simulated networks. squares has to be considered the right one. In all estimated networks internal order of observations and maybe caused by a sequential
While both packages claim to use something like v T Pv → min, which is the weighted sum of squares of the residuals was larger for the SA estimation of network points in SA.
the least square method developed by Gauss, there are siginificant solution than for PANDA. It is planned to analyse more network configurations in the future.
differences in the results. People from other laboratories with different software packages are
welcome to contribute.